What is the Virtual Cell Challenge?
Written by David Kim-Shoemaker
Edited by Rathusha Nimalan
Machine learning has been all the craze in STEM for years now, with it being nearly 10 years since the advent of the Transformer architecture, the first mechanism for encoding attention to specific pieces of information. Originally built for language interpretation, it has since played a pivotal role in the creation of many biological tools and discoveries, such as AlphaFold (a model that outputs a predicted protein sequence) and AlphaGenome (another model from Deepmind that predicts a variety of genetic factors related to genome sequences).
The Virtual Cell Challenge, hosted by the Arc Institute, is just one more application of machine learning that has resulted from the incredible pace the field has been progressing. But what is the Virtual Cell Challenge, and why is it important?
To understand the motivation behind this challenge, a different intuition for cells is required than what is usually considered by the biologist.
Using the framework of the famous short story, “Flatland” by Edwin Abbot, imagine a square: it understands length and width, but height is unimaginable because it exists in a two dimensional world. Because we know what ‘space’ the square lives on and can manipulate it, we can theorize about all kinds of shapes and their properties.
But how can we consider cells? Cells exist in three-dimensional space like we do, their properties depend mostly on the genes they express. Thus, the lens of the computational biologist is in high dimensions: for each gene, a new dimension is added. Instead of plotting a point in 3D, computational biologists plot cells in tens of thousands of dimensions.
So now that we know where cells live (high dimensional space, where each gene is a dimension), it is reasonable to start drawing high dimensional shapes and structures. Unlike the flat paper that was earlier described, there are limits to genes when considering what can exist in theory, and in practice. This is where the Virtual Cell Challenge comes in: Is it possible to determine where cells exist in high dimensions, and the perturbations that cause them to shift into different spaces, without ever testing them on an actual living cell? [1]
To explore this question, scientists employed a very specific tool: Perturb-seq.
Perturb-seq is precisely as it sounds. Identical cells are infected with thousands of different lentiviruses, each containing different guide RNAs. These RNAs knockout specific genes, ‘perturbing’ the cell. After the cells are ‘perturbed’, their gene expressions are measured via single-cell sequencing [2]. In this way, the impacts of thousands of genes can be measured with a single experiment.
Trained on this data, computational biologists employ machine learning techniques to attempt to approximate the manifold: a lower dimensional ‘surface’ of the high dimensional space that represents every cell that may plausibly exist.
The Virtual Cell Challenge is ongoing. This year, the challenge is zero-shot: submitted models will be tested against completely new data, and attempt to recognize perturbations that the models have never been exposed to. Although there is indeed a prize of 100,000$, split evenly between cash and NVIDIA compute credits), the true benefit is priceless.
David Kim-Shoemaker ’(29) is in the College of Agriculture and Life Sciences. He can be reached at djk323@cornell.edu.
Sources:
Virtual Cell Challenge. (n.d.). Virtual Cell Challenge. https://virtualcellchallenge.org/
Genome-Wide Perturb-Seq. (n.d.). https://gwps.wi.mit.edu/
